Masked diffusion decoder scores on seven benchmarks

This CSV transcribes the complete first results table from the UNCODE paper. It contains 19 model or sampler rows, seven benchmark scores per row and the averages reported in the ACL proceedings.

Compare samplers on the same backbone

Filter model_group to LLaDA-8B-Instruct or LLaDA-1.5-8B before comparing an unmasking rule. Each group includes uniform, confidence, entropy, margin, EB-Sampler, Semi-AR, Fast-dLLM and UNCODE results. Three autoregressive reference models appear in their own group. The dagger column preserves the table’s marking for Semi-AR strategies.

The shot count varies by task. HumanEval and MBPP use zero shots, Countdown uses three, GSM8K and MATH500 use four, and GPQA and Sudoku use five. This author-prepared transcription preserves the published scores, including the reported averages. Table 1 in the ACL paper

Data files

Hugging Face dataset card

Related paper

Empirical Analysis of Decoding Biases in Masked Diffusion Models · Official BibTeX

Browse research data · Download catalog (JSON-LD)